After more discussions elsewhere [1], I noticed another problem with
the dynamic completion table from python.el.

Each call to the completion table requires communication with the
inferior process.  If called frequently enough, the comint input
filters can get confused, and some garbage gets printed to the shell.
If the inferior is running over Tramp, the problem gets amplified.

Now, this is not likely to cause much trouble when calling
`completion-at-point' manually, but it is a problem with any of
continuously-updating completion UIs like Company or Corfu.

To alleviate this, the completion-at-point function could implement some
sort of caching.  The difficult question is when to invalidate the
cache.  I've attached one possiblility as a draft patch.  If the
approach seems reasonable, then I'll format it properly.

As a side effect, the patch also solves the original issue described in
this ticket.

Please let me know what you think.

[1] https://github.com/oantolin/orderless/issues/79